pacman : predicting ac consumption minimizing aggregate ... · total. pacman achieved more than 90%...
TRANSCRIPT
PACMAN : Predicting AC Consumption MinimizingAggregate eNergy Consumption
Student Name: Milan Jain
IIIT-D-MTech-CS-MUC-12-067July 28, 2014
Indraprastha Institute of Information TechnologyNew Delhi
Thesis CommitteeAmarjeet Singh (Chair)
Pushpendra SinghZainul M Charbiwala
Submitted in partial fulfillment of the requirementsfor the Degree of M.Tech. in Computer Science,
with specialization in Mobile and Ubiquitous Computing
c©2014 Milan JainAll rights reserved
Keywords: HVAC, User feedback system, Residential Cooling, Energy optimization, Smartbuildings, and Real World studies
Certificate
This is to certify that the thesis titled “PACMAN: Predicting AC Consumption Mini-mizing Average eNergy Consumption” submitted by Milan Jain for the partial fulfillmentof the requirements for the degree of Master of Technology in Computer Science & Engineeringwith specialization in Mobile and Ubiquitous Computing is a record of the bonafide work carriedout by him under my guidance and supervision in the Mobile and Ubiquitous Computing groupat Indraprastha Institute of Information Technology, Delhi. This work has not been submittedanywhere else for the reward of any other degree.
Dr. Amarjeet SinghIndraprastha Institute of Information Technology, New Delhi
Abstract
Buildings account for a significant proportion of overall energy consumption across the world.Heating Ventilation and Air Conditioning (HVAC) typically consumes a major proportion (e.g.32% in India) of the total building energy consumption. While centralized HVAC systemsare more prevalent in developed countries, separate room level Air Conditioners (ACs) are acommonplace in developing countries, such as India. Poor building insulation in developingcountries, together with an option to easily control room level air conditioning, presents amajor opportunity for energy conservation in these countries. We propose PACMAN - a novelapproach for predicting the energy consumption of room level AC. PACMAN involves learninga thermal model of the room from historical usage and combines this model with the weatherforecast for user’s location to guide the user towards optimized AC settings in order to balanceuser comfort and energy efficiency. Empirical validation was performed using a real worldstudy, conducted across 7 homes in India, with collective data for a duration of 2200 hours intotal. PACMAN achieved more than 90% accuracy in predicting the energy consumption acrossdifferent ACs, room types and set temperatures used during the data collection. We furtherdescribe a prototype realization of the proposed PACMAN system towards achieving reducedAC energy consumption with better feedback and control.
Acknowledgments
I am grateful to my parents for their emotional and moral support which always encouraged mein my Masters. Without them, this journey would not have been possible.
I thank IIIT Delhi faculty members, Dr. Rajiv Raman, Dr. A. V. Subramanyam, Dr. P. B.Sujit, Dr. Vikram Goyal, Dr. Sandip Aine, Dr. Sanjit Krishnan Kaul, Dr. Shreemoy Mishraand our Director Dr. Pankaj Jalote to allow us to perform experiments across their apartmentsand actively participate in data collection. I would also like to thank members of energy group,Nipun Batra, Pandarasamy Arjunan and Manoj Gulati, for their guidance in performing theexperiment, and critique reviews. I would also like to thank Alvika Gautam, Garvita Bajaj andParkshit Maini for being my batchmate and providing me such a supportive environment duringthe Master’s course, which helped me immensely.
Finally and importantly, I would like to thank Dr. Amarjeet Singh for being my mentor andadvisor for the whole master course. Your creativity, work ethics and time management skills,motivate and inspire me to do quality work.
i
Contents
1 Research Aim and Motivation 1
2 Related Work 5
3 Research Contributions 7
4 Background 8
4.1 Thermodynamics of the Room . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
4.2 Room Level Air Conditioners . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2.1 Vapor Compression Cycles . . . . . . . . . . . . . . . . . . . . . . . . . . 9
4.2.2 Working of Air Conditioner . . . . . . . . . . . . . . . . . . . . . . . . . . 10
5 PACMAN 12
5.1 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
5.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.3 System Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
5.3.1 PACMAN-L . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
5.3.2 PACMAN-P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
5.3.3 PACMAN-Engine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
6 Evaluation 22
6.0.4 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
6.0.5 Data Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
6.0.6 Analysis of AC Energy Prediction Accuracy . . . . . . . . . . . . . . . . . 23
7 Discussion 27
8 Realization of PACMAN 29
8.1 Z-Wave Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
8.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
8.2.1 System Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
ii
8.2.2 System Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
8.2.3 User Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
9 Conclusion 33
iii
List of Figures
1.1 Results from the survey of 1800 people in Urban India,discussed in Jain et al. [12],illustrating (a) perceptions about AC power consumption and (b) opinion on scopefor reduction. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Temperature and power consumption for AC usages with different set temperatureand external temperature. Higher set temperature results in significant savingswith compressor being turned off for most of the time. . . . . . . . . . . . . . . . 3
1.3 Per hour energy prediction at different AC set temperatures across all usagesobserved during our study. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Illustration of Nonintrusive Load Monitoring1. Using NILM to disaggregate en-ergy consumption by Refrigerator, Microwave and Light from total power con-sumption. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
4.2 Variation at different Set Point Temperature . . . . . . . . . . . . . . . . . . . . 11
5.1 Use case of PACMAN: When the AC is switched ON, the set temperature andweather forecast is used with the learned model to predict the AC power con-sumption. When the AC is switched off, the room temperature data, collectedduring the past usage, and corresponding external temperature data is used toimprove the learning process. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
5.2 PACMAN System Architecture illustrating interaction of different sub-systems,within PACMAN system, with each other. . . . . . . . . . . . . . . . . . . . . . . 14
5.3 Variation of room temperature with AC compressor cycles (set temperature ofAC being 27◦C) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
6.1 Installed Raspberry Pi for Data Collection . . . . . . . . . . . . . . . . . . . . . . 22
6.2 Validation of different datasets used for the empirical evaluation of PACMAN. . 23
6.3 Variation in compressor on and off temperatures across different set temperaturesand usages. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
6.4 Comparison of AC Energy Prediction accuracy based on different influencing pa-rameters - room (for thermal properties), set temperature (being directly used inthe prediction model) and AC manufacturer (possibly leading to different controlmechanisms). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
6.5 Performance evaluation of PACMAN based on (a) Status prediction accuracy (b)Precision (c) Recall. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
iv
7.1 Illustration of outlier usages resulting in reduced AC energy prediction accuracy 28
8.1 System Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
8.2 User Interface to monitor AC power consumption and control AC remotely . . . 32
v
List of Tables
5.1 Summary for each room from where data was collected. Second column representsnumber of days we collected the data, third column shows various set tempera-tures used during the study. AC technical specifications are as supplied by themanufacturer. Room specifications, that impact thermal modeling, includes oc-cupancy denoting number of occupants for the room and floor ratio representingproportion of room’s floor number to total floors in the building. . . . . . . . . . 20
vi
Chapter 1
Research Aim and Motivation
Energy crisis is a major issue across the globe. Governments, around the world, are concerned
about energy crisis, as electricity demand is rising at an alarming rate. Of this total electricity
demand, International Energy Agency (IEA) estimated the contribution of buildings to be one-
third1. Buildings also contribute to CO2 emissions which is one of the major sources of pollution
in the environment1.
While energy crisis is a global issue, conditions are even worse in developing countries, such as
India, when compared to developed countries. Even being the third largest electricity producer,
25% of the population in India have no access to electricity at all along with the shortage of
9% during peak demands2. Of all the electrical appliances, Heating, Ventilation and Air Condi-
tioning (HVAC) contributes a significant proportion (30-50%) of the total energy consumption
across both residential and commercial buildings3,4,5 [16]. However, most of the developing
countries have a higher prevalence of room level ACs in residential buildings. Possibility of
room level control, together with building’s substandard thermal insulation make decentralized
ACs, an attractive target for energy conservation in residential buildings, which consume signif-
icant amount of electricity. According to a previous study carried out globally in 2005 by IEA,
residential buildings contribute 29% in aggregate energy consumption across the globe6. They
also contribute 21% directly and indirectly to global CO2 emission6. While the savings at an
individual home level may seem insignificant but, when scaled up to millions of homes across
the globe, will help us to face challenging energy demands.
On the other side, Centralized air conditioners (AC) in a residential building involves division
of conditioned space into multiple zones, each consisting of multiple rooms serviced together
by independent fan units. Difference in occupancy levels, together with higher possibilities for
leakage of cooling, across rooms within a zone make optimal control for a central AC system
1http://bit.ly/1m2mXyG2http://bit.ly/1oWwuHz3http://bit.ly/1icjKIJ4http://bit.ly/1ow8fjp5http://bit.ly/1ol1tv66http://bit.ly/1mhY4yC
1
(a) Power consumption estimate (b) Expected usage
Figure 1.1: Results from the survey of 1800 people in Urban India,discussed in Jain et al. [12], illustrating(a) perceptions about AC power consumption and (b) opinion on scope for reduction.
difficult.
However, lack of understanding among the users to optimize the energy consumed during an AC
usage (Time duration from the time AC is switched on to the time it is switched off), results
in significant wastage. A survey conducted across 1800 people in urban India [12] observed two
interesting facets of this energy wastage. First, many of the users (more than 76% of those
surveyed) underestimate the power consumption of ACs (see Figure 1.1a for detailed breakup).
Secondly, majority of users (more than 50% of those surveyed) believe that there is a scope
to save energy during their AC usage when compared to their current usage (see Figure 1.1b
for detailed breakup). Previous study [15] has observed that occupants, who are unaware of
their electrical consumption, add up to one-third to their electricity bills. However, real time
and appliance level feedback about energy consumption together with steps on how to achieve
energy efficiency is largely missing today.
An interesting extension of current commercial AC control systems is a feedback system that
guides the user to achieve the desired comfort level while still being able to attain energy
efficiency. Specifically for AC, such a feedback system should accurately predict the energy con-
sumed as per the user’s desired set temperature and inform her on possibilities of energy savings
by increasing the set temperature by a few degrees. It is important for prediction to be accurate
so that the user trusts the system. Thus for accurate prediction, the system should account for
outside weather conditions together with the learned thermal model from the historical usage.
Figure 1.3 presents AC energy consumption per hour at various set temperatures (25-29)◦C used
across all the ACs during our data collection. Ideally, energy consumption should increase with
a decrease in the set temperature. However, as can be observed, it varies depending upon the
thermal environment in the room, and external weather conditions. Thermal environment for a
room in turn needs to account for quality of insulation, occupancy count and human activities
being undertaken. Correspondingly, accurate prediction of AC energy consumption requires a
2
(a) 16◦C set temperature with average externaltemperature at 28 ◦C.
(b) 28◦C set temperatuer with average externaltemperature at 28 ◦C.
Figure 1.2: Temperature and power consumption for AC usages with different set temperature andexternal temperature. Higher set temperature results in significant savings with compressor being turnedoff for most of the time.
thermal model of the room environment, together with accounting for AC settings and external
weather conditions.
In this work, we focus on real time feedback about AC usage in order to optimize the energy
consumption. Commercial products for optimization of AC energy consumption, such as Nest7,
learn user settings over a period and subsequently control AC to reduce energy wastage. How-
ever, such a system lacks feedback to the user, about the impact of their current settings on the
energy consumption, to potentially motivate change in behavior for setting up slightly higher
set temperatures. As an illustration of possible energy savings with varying set point tempera-
tures, Figure-1.2 presents room temperature and power consumption at different set points and
external weather conditions as collected during real usages within different homes during our
data collection. Figure-1.2a present the data with set point at 16◦C whereby AC compressor
remained ON all throughout the usage from 7:25 - 8:40 PM. However, when the set temperature
was 28◦C, it was quickly achieved in the desired room, and thereafter AC fans were enough to
maintain the temperature, allowing the AC compressor to remain off. Overall, with set point
at 28◦C, a total of 20 units of electricity were saved over the duration of 10 hours of operation
illustrated in Figure-1.2b.
Motivated from the energy saving opportunities, presented by the large number of room level
ACs being used in India, we propose PACMAN, to forecast AC energy consumption based
on its set temperature and weather forecast for the user’s location. PACMAN estimates AC
energy consumption without using any power measurements, by accounting for variation in
room temperature during historical usage and the rated power consumption of the AC. Such
temperature measurements can be easily obtained from a room level sensor or even with mobile
7https://nest.com/
3
Figure 1.3: Per hour energy prediction at different AC set temperatures across all usages observed duringour study.
phones that come with embedded temperature sensors (such as Samsung Galaxy S48 and Galaxy
Note 39). PACMAN provides access to the real-time room temperature (as communicated by
the sensor), weather forecast for the user location and current external temperature (pulled
from a weather website) and predicts energy consumption around the desired set temperature
as a feedback to motivate the user towards optimally setting the AC temperature. PACMAN
updates its model parameters based on each historical usage of the AC, thus providing improved
performance with time. PACMAN is evaluated using real data (consisting of both room level
temperature and AC power consumption) collected from 6 different ACs used across 9 rooms in
7 homes, for a total duration of 2200 hours.
Rest of this thesis is organised as follows. In Chapter-2, we provide an overview of related work.
Chapter-5 exhibits a use case scenario of PACMAN to understand its role in energy conserva-
tion followed by PACMAN system and its components and concluded with the implementation
details of PACMAN. Chapter-6 evaluates its performance using data from a real world study.
In Chapter-7, we discuss various outliers, limitations and future directions for the PACMAN
system. Finally, Chapter-8 provides details of prototype realization of PACMAN system, before
we conclude in Chapter-9.
8http://global.samsungtomorrow.com/?p=236109http://bit.ly/1qb5fsi
4
Chapter 2
Related Work
In this chapter, we present various approaches to provide disaggregated power consumption to
the user. We also look into existing techniques to optimize energy consumption by HVAC and
personal air conditioners.
Optimization of HVAC systems towards energy conservation has been well studied in the past.
Much of this optimization work is done on central HVAC systems, optimizing their operations
based on the occupancy and outside weather conditions. Previous studies [1, 8, 9, 14] have used
passive infrared sensor (PIR) to detect occupancy and sleeping patterns within the area to control
the HVAC systems. Agarwal et al. [2] observed PIR sensors to be inadequate for HVAC control
due to error in detecting relatively stationary occupants. Besides direct occupancy monitoring
using PIR sensors, indirect sensing techniques such as detecting changes in the concentration of
CO2 levels1,2, and using logs of wireless access points [3] have also been used to detect occupancy
for controlling the centralized HVAC system. However, this line of work does not involve the
occupant in the loop for deciding the optimal AC settings. Participatory involvement of occupant
to decide optimal AC operations is a core concept, embedded in the proposed PACMAN system.
For decentralized HVAC systems, more commonly used in residential settings, Scott et al. [20]
proposed sensors based actuation while Rogers et al. [19] introduced a system to control heating
based on calculated thermal comfort in real-time. An automated sensor based HVAC control
could have false positives from sensors leading to energy wastage or false negatives leading to
user discomfort [4]. Involving the occupant, to decide the optimal control, can further help to
balance the user comfort and energy efficiency appropriately.
Several studies have shown the impact of providing the feedback to the users on the energy
conservation. Detailed feedback of appliance level energy consumption was observed to result in
5-15% savings in energy consumption [7]. Such an appliance level feedback could be provided by
directly monitoring and controlling the plug loads [21] or using a single meter at the mains level.
1http://bit.ly/1mtS7An2http://bit.ly/1rxFwhk
5
Figure 2.1: Illustration of Nonintrusive Load Monitoring5. Using NILM to disaggregate energy consump-tion by Refrigerator, Microwave and Light from total power consumption.
Main meter data is disaggregated into appliance level consumption3,4 [11] using machine learning
algorithms. Figure-2.1 shows an illustration of Nonintrusive load monitoring where authors
developed a Non-Intrusive Load Monitoring Tool Kit [5] to detect refrigerator, microwave and
light from aggregate power consumption of the house. PACMAN predicts energy consumption
of AC without directly monitoring the energy usage. It leverages the temperature sensor from
the room, and outside temperature extracted from a weather website, thus making the overall
system both inexpensive and easier to maintain.
Recent work from Rogers et al. [18] is the closest to our work. They proposed MyJuolo system
whereby a USB based temperature logger collected room temperature data that was used to learn
a thermal model for the room, involving several parameters including leakage rate, heating rate
and thermal noise within the room. However, their process requires significant user intervention
and involves learning a static thermal model that may not provide accurate predictions with
varying operating and weather conditions. In contrast, PACMAN requires minimal effort from
the user and updates the learned model with each AC usage, thus optimally accounts for dynamic
operating and weather conditions.
3https://plotwatt.com/4https://www.bidgely.com/5https://github.com/nilmtk/nilmtk
6
Chapter 3
Research Contributions
In this section, we will discuss about novel contributions of our work in the context of existing
literature on heating, ventilation and air conditioning(HVAC) systems. Major contributions of
this work are:
1. We propose a feedback system PACMAN, Predicting AC Consumption Minimizing Ag-
gregate eNergy Consumption, to forecast energy consumption by AC based on its set
temperature, and weather forecast for user’s location.
2. We present architecture and implementation details of the system, used to collect room
temperature data. We also collected meter data and weather information used in evalua-
tion of PACMAN’s performance.
3. We analyse reasons for some of the outliers in prediction accuracy of PACMAN, caused
primarily from imperfect applicability of the proposed thermal model, thus requiring in-
clusion of multiple thermodynamic properties to enrich the current model being learned.
4. Finally, we explain a prototype system realization using PACMAN to provide feedback to
the user towards optimal usage of AC.
7
Chapter 4
Background
In this section, we discuss the thermodynamics involved during the interaction between room and
AC. Then we look into the fundamental concept of air conditioner i.e. Vapor Compression Cycle,
widely used method for air conditioning. Finally, we present working of an air conditioning unit
and its various components.
4.1 Thermodynamics of the Room
Li et. al. [13] proposed a model of the Air Conditioning and Refrigeration (AC&R) system used
for temperature regulation in enclosed space. They considered room as a system interacting
with another thermodynamic system i.e. AC&R. AC&R extracts room air, works (cooling) over
it and supply back to the room at desired temperature.
However, our thermal model considers room to be an open space interacting with vicinity around
it. Figure 4.1a presents the thermal model for the heat exchange between room, air conditioner,
and the surrounding. AC extracts room air at higher temperature, absorbs heat from the air and
supplies it back to the room at user desired temperature. AC throws, heat extracted from the air
to the vicinity outside the room. AC is a machine which takes room air at higher temperature
and provides air to the room at desired temperature plus ejects heat outside the room. The rate
at which air conditioner absorbs heat from room air is termed as Cooling Rate. Cooling Rate
defines the performance of the AC, as lower the cooling rate more the time it will take to cool
down the room.
When AC is on, room temperature goes down and to maintain thermal equilibrium between
inside and outside room temperature, heat gets added in the room from outside. Source could
be windows, walls, open doors, etc. The rate at which heat gets added in the room due to the
structure of the building is termed as Leakage Rate. Higher leakage rate affects the thermal
efficiency of AC as chillness added to the room, leaks to the surrounding environment. Better
thermal insulation of the room helps in reducing leakage rate.
Heat transfers through conduction, convection and radiation. Thus, temperature varies because
8
(a) Thermodynamics of the Room (b) Vapour Compression System[17]
(c) Schematic diagram of a RoomLevel Air Conditioner
of multiple sources, such as number of occupants, human activities, that radiate or absorb heat
in the room. These sources are random in nature and varies with time. Variation in room
temperature due to random thermal sources is termed as Thermal Noise. It creates interference
in the working of AC and effect of thermal sources depend on their nature.
4.2 Room Level Air Conditioners
In this section, we discuss the working of a room level air conditioner and its various components.
We start from vapor compression cycle, used in ACs.
4.2.1 Vapor Compression Cycles
AC leverages the concept of vapour compression refrigeration cycle which completes four fun-
damental mechanical processes; (1) Compression, (2) Condensation, (3) Expansion, and (4)
Vapourisation. In this section we introduce important terms used in AC and major components
in vapour compression system:
• Referigerant: Rajput et. al. [17] defined Refrigerant as “Any substance that absorbs
heat through expansion or vaporisation and loses it through condensation in a refrigeration
system. These substances absorb heat at one place at low temperature level and reject the
same at some other place having higher temperature and pressure. The rejection of heat
takes place at the cost of some mechanical work.”
• Compressor: Its function is to remove refrigerant’s vapours from evaporator and increase
the temperature and pressure of refrigerants to a point such that its vapours can be
condensed by the condensor.
• Condensor: It provides a medium to refrigerant vapours, to pass heat and liquify itself.
• Expansion Valve: It controls amount of refrigerant going to the evaporator and reduce
pressure its pressure so that liquid will vapourize in the evaporator at the desired low
temperature and take out sufficient amount of heat.
9
• Evaporator: It provides a heat exchange medium to transfer heat from room air to
refrigerant vapours.
Figure-4.1b presents a vapour compression system. Refrigerant circulates in the system and
switches it state from liquid to vapour, vapour to liquid back and forth. Following are the steps
invloved in the cycle:
1. Low pressure refrigerant vapours enter the compressor, where it is compressed to increase
its pressure.
2. High temperature and high pressure vapours enter the condensor where they are condensed
to high pressure liquid referigerant and collected in a receiver tank.
3. From receiver tank, refrigerant vapours pass through expansion valve where they are throt-
tled down to a low pressure and low temperature liquid referigerant.
4. Finally vapours are passed on to the evaporator where it extracts heat from the surrounding
and changes to low pressure refrigerant vapours.
4.2.2 Working of Air Conditioner
In this section, we discuss working of a room level air conditioner and its various components.
Figure 4.1c presents the schematic diagram of an air conditioning unit. Two fans are present
for air circulation, one towards evaporator coils and other towards condenser coils. Fan towards
evaporator coils draws air from the room and then it is passed over the evaporator coils containing
low pressure low temperature liquid refrigerant. It absorbs heat from the room air and supplies
cooler air to the room. Due to heat absorption, refrigerant liquid converts to high temperature
refrigerant vapours. These vapours are passed to the compressor for increasing the pressure and
temperature for condensation.
Fan towards condenser coils extracts air from the outside environment at high temperature.
Extracted air is passed to condenser coils having high pressure and high temperature refrigerant
vapours. Air absorbs heat from refrigerant and fan circulates it back to the outside environment
at higher temperature thus converting refrigerant vapours to low temperature liquid refriger-
ant and passed to evaporator coils. Amount of refrigerant passing through evaporator coils is
controlled by the expansion valve using the Set Temperature of the AC, which is the room tem-
perature desired by the user. This cycle continues till the thermostat senses room temperature,
close to the user desired temperature.
A single vapour compression cycle is termed as Compressor Cycle. Compressor consumes the
maximum power consumption for AC. Power consumption by a compressor depends upon the
set temperature. As we decrease the set temperature, volume of refrigerant vapours going into
the compressor also increases thus increasing the compressor’s power consumption. AC Usage is
the time user switches on the AC unit, till she switches it off. A single AC usage may comprise
of one or more compressor cycles.
10
(a) Set Point Temperature at 20◦C (b) Set Point Temperature at 26◦C
Figure 4.2: Variation at different Set Point Temperature
According to Clausius statement, ”Heat can never pass from a colder to a warmer body without
some other change, connected therewith, occurring at the same time” [6]. Therefore, we need to
do some work to extract heat out of the room to the surrounding. Air Conditioner does this job
and work done results in energy consumption. Lower the temperature you set in air conditioner,
more energy it would consume. Figure 4.2 shows variation in energy consumption based on set
point temperature. During our study we found, some temperatures set points were infeasible
according to that environment. Room was not able to attain that specific set point temperature
because of which compressor remain turned on for whole duration air conditioner was on. This
results in higher energy consumption by air conditioner. We will discuss in next section about
the savings PACMAN can do by predicting energy consumption based on set temperature and
provide updates on what minimum temperature could be achieved in specified time duration at
this set point.
11
Chapter 5
PACMAN
In this chapter, we discuss the design, architecture and implementation details of PACMAN.
First we provide an overview of PACMAN through a use case scenario followed by discussion
about the architecture and work flow of PACMAN and conclude with the implementation details
of PACMAN.
5.1 System Overview
Before getting into the details of PACMAN system, we discuss a use case that envisions it
in a real-world setting. Figure 5.1 illustrates the usage scenario. Alice uses PACMAN web
interface to input the temperature at which she desires to operate her AC. PACMAN pulls the
predicted weather conditions, for the area surrounding Alice’s residence, from a weather service.
Weather forecast, together with the thermal model of Alice’s room is used to predict the energy
consumption around the desired AC set temperature.
Alice also has a temperature sensor in her room that is uploading temperature data to a server
accessible by PACMAN. This server could be one of the many Internet of Things (IoT) ser-
vices available today (such as Xively1) and the temperature sensor could be one of the many
off the shelf available in the market (e.g. Maxim’s 1-Wire DS18B20 temperature sensor2) or
temperature sensor in Alice’s smartphone (Samsung Galaxy S48, Galaxy Note 31). As the AC
is switched off, PACMAN pulls the observed room temperature from the server and outside
weather conditions from the weather service, corresponding to the duration when AC was on,
to update the thermal model of the room.
Real time room temperature together with current and predicted outside weather conditions
are also shown by PACMAN to Alice for better user engagement. Besides the web interface,
PACMAN also comes as a smartphone application for easy accessibility by Alice.
1www.xively.com2http://bit.ly/1fdT62x
12
26
28
30
Te
mp
era
ture
(°C
)
20
25
30
35
Te
mp
era
ture
(°C
)
20
25
30
35
Tem
pe
ratu
re (
°C)
Figure 5.1: Use case of PACMAN: When the AC is switched ON, the set temperature and weatherforecast is used with the learned model to predict the AC power consumption. When the AC is switchedoff, the room temperature data, collected during the past usage, and corresponding external temperaturedata is used to improve the learning process.
5.2 System Architecture
We designed PACMAN to forecast energy consumption of AC for the user at different set
temperatures, inspiring them to be energy conscious. Figure 5.2 presents the architecture of
PACMAN, illustrating its various components and interaction between them. Keeping in mind
the modularity and extensibility of the system, PACMAN is divided into four major sub-systems;
1. PACMAN-L is a learning component that utilizes set temperature, room and outside
temperature, during the AC usage, to learn the following thermal parameters for the
room:
(a) Cooling Rate: The rate at which AC removes heat from the room.
(b) Leakage Rate: The rate at which heat gets added in the room because of its ther-
mal insulation and structure. Structure includes various factors such as number of
windows, area of the room and material used for the walls.
(c) Thermal Noise: Room temperature could also vary due to factors such as occu-
pancy, human activities in the room and heat generated by various appliances in the
room. Thermal noise indicates the effect of these sources on the room temperature.
2. PACMAN-P is the prediction component providing energy consumption at different set
temperatures. When the AC is switched on, PACMAN-P takes as input the set tempera-
ture, thermal parameters learned from previous usages and predicted outside weather to
then forecast the energy consumption of AC for the current usage.
3. PACMAN-UI takes set temperature and user’s location as input to generate energy fore-
cast for AC along with the weather forecast information for that location. Further, it also
displays room temperature and external temperature in real time. When user switches off
the AC, PACMAN-UI also communicates energy consumption for that usage to the user.
13
Figure 5.2: PACMAN System Architecture illustrating interaction of different sub-systems, within PAC-MAN system, with each other.
4. PACMAN-Engine is an administrative component that controls other components of PAC-
MAN, as shown in Figure 5.2. It interacts with a weather service to get historical weather
data to learn thermal model of the room, and weather forecast for prediction of energy
consumption. It also interacts with the IoT server to get room temperature data.
PACMAN-Database manages various records that are used to cache statistics generated by all
other components. Thermal parameters learned by PACMAN-L are accumulated in L-Records
while P-Records manage prediction statistics such as energy forecast and thermal parameters
applied for prediction. LP-Statistics stores analysis, related to accuracy metrics such as pre-
diction error, to enhance the overall performance of PACMAN. PACMAN-Engine is the only
system component with the authority to access and manipulate the database.
5.3 System Implementation
PACMAN is implemented in Python while utilizing several libraries supporting machine learn-
ing tools (e.g. Pandas and scikit-learn). We now discuss the implementation details for each
component of PACMAN, as outlined in Section 5.2.
14
5.3.1 PACMAN-L
It is the learning component of PACMAN that takes set temperature, collected room tempera-
ture data and weather conditions during the AC usage, to learn different thermal parameters i.e.
(1) Cooling Rate, (2) Leakage Rate, and (3) Thermal Noise. This component includes multiple
modules as explained next.
Usage Parser
After switching off the AC, a usage matrix is generated constituting time annotated room, and
outside temperature data measured at mismatched sampling rate. To time align the collected
data, this module alters their sampling rate to 1Hz and interpolates the missing value using
linear interpolation3. Thereafter, it segregates timestamp, room temperature and outside tem-
perature vectors from the usage matrix. Finally, it estimates the duration of AC usage using
the timestamp and room temperature vector.
Status Vector Generator
Usage parser module generates a room temperature vector which defines an AC usage, compris-
ing of one or more compressor cycles. We assumed AC to be a two state machine; (1) Compressor
On, and (2) Compressor Off. For every usage, this module generates a status vector which is
an array of binary numbers (0,1) to present the compressor state at any moment during the AC
usage. Status vector, using Equation-5.1, estimates energy consumption for an AC usage where,
Prated is the rated power consumption of AC as specified by the manufacturer.
Econs = Prated ×∑{[1, 1, 0, ...0, 0]} (5.1)
Power Temperature Correlation: We use temperature data to depict AC state at any
instant. Figure 5.3 exhibits the correlation between the room temperature and AC power con-
sumption. As the set temperature (27◦C) is achieved in the room, the compressor is turned off
resulting in an increase in the room temperature. When the room temperature goes beyond
a threshold (this threshold varies for different AC), the compressor is again turned on, and
the room temperature starts decreasing. Correspondingly, room temperature keeps fluctuat-
ing around the set temperature with minimum corresponding to compressor off instance and
maximum corresponding to compressor on instance.
Compressor On & Off Temperature learner
AC thermostat temperature, resulting in switching on and off of AC compressor during a usage
depends upon AC’s control algorithm. However, room temperature, as observed at another
3http://bit.ly/TJwM9j
15
Figure 5.3: Variation of room temperature with AC compressor cycles (set temperature of AC being27◦C)
location in a room, depends upon numerous factors such as number of occupants, activity
performed by the occupants, number of windows and doors in the room and material of walls.
As a result, the temperature recorded by the sensor deployed away from the AC, corresponding
to compressor on or off instance might differ from what the AC thermostat, on which the AC
control cycle depends, might observe.
δon = Ts −n∑
i=0
T ion
n(5.2) δoff = Ts −
n∑i=0
T ioff
n(5.3)
To learn the correlation between temperature observed by the room sensor and the state of AC
compressor, we use the temperature observed during the compressor on and off instances for each
cycle from the room temperature vector. As threshold temperature value depends upon control
algorithm of AC, δon i.e. the deviation of compressor on temperature from threshold temperature
value and δoff i.e. the deviation of compressor off temperature from the set temperature are
modeled using AC set temperature as shown in Equation-5.2 and Equation-5.3 respectively. Ts
is AC set temperature, n is the number of compressor cycles and T ion, T i
off are compressor on
and compressor off temperatures respectively for ith compressor cycle.
Thermal Model
While the room temperature at any instant depends upon numerous factors, we explicitly con-
sidered the effect of AC cooling and external temperature for our thermal model. We treated
16
Algorithm 1: Algorithm to Predict AC Compressor Cycles
Input: T0, Ts, tduration, θOutput: TPred, SPred
T(0)Pred = T0 ;
for i = 0 to tduration doT onpred = Ts − δon ;
T offpred = Ts − δoff ;
T(i+1)Pred − T
(i)Pred = α× (T
(i)Pred − T
(i)ext) + β × S(i) + ε ;
if T(i+1)Pred ≤ T
(i)Pred then
if T(i+1)Pred ≥ T
OffPred then
S(i+1)Pred = 1 ;
else
S(i+1)Pred = 0 ;
end
else
if T(i+1)Pred ≤ T
OnPred then
S(i+1)Pred = 0 ;
else
S(i+1)Pred = 1 ;
end
end
end
all other factors as thermal noise in the room.
T(i+1)int − T (i)
int = α× (T(i)int − T
(i)ext) + β × S(i) + ε (5.4)
Equation-5.4 presents PACMAN’s thermal model using status vector, room and outside tem-
perature and thermal parameters of the model i.e. (1) Cooling Rate (β), (2) Leakage Rate (α),
and (3) Thermal noise (ε). T(i)int is the room temperature, T
(i)ext presents external temperature,
and S(i) depicts AC state, at ith second during the AC usage.
PACMAN generates a feature matrix from room temperature vector and external temperature
vector generated by usage parser module. It then computes temperature gradient for every
second using room temperature vector to generate output matrix. Thermal parameters corre-
sponding to the usage are learned using linear regression over feature and output matrix.
5.3.2 PACMAN-P
The prediction component of PACMAN captures AC set temperature from the user, together
with her location, to forecast AC energy consumption. Using the learned temperature deviations
for compressor switching, as explained in Section 5.3.1, this segment predicts the compressor on
and off temperature using Equation-5.5 and Equation-5.6 respectively.
17
T onpred = Ts − δon (5.5) T off
pred = Ts − δoff (5.6)
We present prediction algorithm (Algorithm-1) to predict status vector for AC, which estimates
energy consumption of AC using Equation-5.1.
The algorithm uses optimal thermal parameters, θ (={α, β, ε, δon, δoff}), current room temper-
ature (T0), AC set temperature and weather forecast, to predict room temperature and status
vectors for the entire usage (tduration). T(i)pred denotes predicted room temperature and S
(i)pred is
the predicted compressor state at ith second. T 0pred denotes room temperature at the time AC
was switched on. PACMAN-Engine selects optimal thermal parameters from L-Records using
optimal parameter selection algorithm illustrated in Algorithm-2.
5.3.3 PACMAN-Engine
We now discuss the administrative component of PACMAN that manages the learning and the
prediction sub-systems. It is the only component in the whole framework which is authorized
to access the internal database of PACMAN and interact with external weather or temperature
server. PACMAN-Engine uses a weather service for weather forecast, historical weather data
and real time update of outside temperature.
Accuracy Model
In order to evaluate prediction, PACMAN-Engine provides a set of metrics. We now give a brief
description of each metric implemented in PACMAN along with its mathematical definition.
1. Prediction Error and Accuracy in Energy Consumption: We evaluate PACMAN perfor-
mance through prediction error (Eerrorpred ) and prediction accuracy (Eacc
pred) of AC energy
consumption. Prediction error is defined as:
Eerrorpred =
Epred. − Econs
Econs× 100 (5.7)
while, prediction accuracy as:
Eaccpred = 100− |Eerror
pred | (5.8)
where, Econs is the energy consumed by AC and Epred is the energy forecast for the same
AC usage. PACMAN-Engine reports prediction error and accuracy to the user, in terms
of percentage and stores them in LP-Statistics.
2. Mean Absolute Deviation(MAD) of Temperature and Status Prediction: Alignment of tem-
perature and status prediction curve from the actual values is defined by Equation-5.9 and
18
5.10 respectively.
Tmadpred =
∑ni=0 |T i
pred − T iact|
n(5.9)
Smadpred =
∑ni=0 |Si
pred − Siact|
n(5.10)
3. True Positives, False Positives, False Negatives, True Negatives: True positives are the
number of time slices in which air conditioner’s compressor was correctly classified as being
on, false positives are number of slices classified as being on while it was off.
TruePositive(TP ) =∑t
AND(Sipred = 1, Si
act = 1) (5.11)
FalsePositive(FP ) =∑t
AND(Sipred = 1, Si
act = 0) (5.12)
TrueNegative(TN) =∑t
AND(Sipred = 0, Si
act = 0) (5.13)
FalseNegative(FN) =∑t
AND(Sipred = 0, Si
act = 1) (5.14)
False negatives are the number of slices where it was classified as off while it was actually
on and true negative are the number of slices when it predicted correctly as off.
4. Precision, Recall: Precision defines the fraction of number of time slices air conditioner’s
compressor was correctly predicted to be on to number of time slices compressor was
predicted to be on. While recall is the fraction of number of time slices air conditioner’s
compressor was correctly predicted to be on to number of time slices compressor was
actually on.
Precision =TP
TP + FP(5.15)
Recall =TP
TP + FN(5.16)
5. F-Score: The harmonic mean of precision and recall.
F − Score =2× Precision×RecallPrecision+Recall
(5.17)
Thermal Parameter Selection
To forecast energy consumption of AC, prediction algorithm (Algorithm-1) need thermal pa-
rameters as one of the inputs. PACMAN-Engine stores thermal parameters learned from each
AC usage, in L-Records of PACMAN-Database. L-Records, together with thermal parame-
19
Experimental Setup Air Conditioner Technical Specifications Room Specifications#ID #Days Used Set Tem-
peraturesModel Tonnage Rated
Power(Watts)
FloorArea(sq.mt)
WindowWidth(meter)
Occupancy FloorRatio
1 15 26, 27, 28 O-GeneralAXZB18GNL-W
1.5 2180 20.06 2 1-2 2/2
2 9 26 O-General 1.5 2180 13.2 2 2-3 2/123 9 25, 26, 27, 29 O-General 1.5 2180 14.5 2 1-2 2/124 13 25, 26, 27 Hitachi RAU518HTD 1.5 1570 20.06 2 1-2 5/125 41 26, 27, 28 Haier HW-18L2H/2013 1.5 1950 20.06 2 1-2 5/126 9 25, 26, 27 Carrier Durakool
GWRAC 018DR0021.5 1950 13.2 2 3-4 6/12
7 12 26, 27, 28, 29 LG L-Crescent 1.5TR5Star
1.5 1600 20.06 2 1-2 7/12
8 9 16, 18, 20, 26,28, 29, 30
Hitachi RAV518HTD 1.5 1575 20.06 2 1-2 9/12
9 10 18, 26, 27 Voltas 4501468/2012Platinum
2 1984 42.2 3.5 1-2 9/12
Table 5.1: Summary for each room from where data was collected. Second column represents numberof days we collected the data, third column shows various set temperatures used during the study. ACtechnical specifications are as supplied by the manufacturer. Room specifications, that impact thermalmodeling, includes occupancy denoting number of occupants for the room and floor ratio representingproportion of room’s floor number to total floors in the building.
ters, contain one more column which is termed as score. After every prediction, we evaluate
PACMAN performance based on accuracy in energy forecast. Based on the achieved accuracy,
PACMAN-Engine updates the previous score of the row in L-Records using Equation-5.18.
score = score+
+2, if 90 < Eaccpred ≤ 100
+1, if 80 < Eaccpred ≤ 90
+0, if 70 < Eaccpred ≤ 80
−1, if 50 < Eaccpred ≤ 70
−2, if otherwise
(5.18)
Score of a thermal parameter symbolizes its performance when used for energy forecast. With ev-
ery newly learned usage, a score of zero is initialized to thermal parameters. Further, PACMAN-
Engine uses Algorithm-2 to select optimal thermal parameters for AC energy prediction. The
algorithm selects five top scores having the same phase and the same set temperature from the
L-Records. Phase depicts whether current room temperature is less than or more than average
external temperature forecast. A set of room temperature and external temperature, Econd,
presents thermal environment of the room. We then calculate euclidean distance (dist) between
every record from the selected L-Records and current room and external temperature. Mini-
mum. amongst the calculated euclidean distances, represents the closest thermal environment,
from the historical usages, for the current AC usage. Eprevcond is an array of a set {Tint, Text} for
each of the selected L-Records. Ecurrcond depicts current environmental condition and IEDist is an
array having euclidean distance of each parameter from Ecurrcond.
20
Algorithm 2: Algorithm to Select Optimal Thermal Parameters
Input: pValue, Ts, Tmeanext , T 0
int
Output: θlrecords = select top five score from L-Record where phase = pValue and Tset = Ts ;Eprev
cond = select T iint, T
meanext from each record in lrecords;
Ecurrcond = {T 0
int, Tmeanext } ;
for i in Eprevcond do
dist = Euclidean distance{Eprevcond[i], Ecurr
cond};IEDist = IEDist.add(dist)
endindex = indexOfMin(IEDist);θ = lrecords[index];
Record Update
PACMAN-Engine maintains a single database that comprises of three tables; (1) L-Records, (2)
P-Records, and (3) LP-Statistics to cache intermediate information by various segments. We
implemented PACMAN database in MySQL.
L-Records maintain information such as leakage rate (α), cooling rate (β), thermal noise (ε),
compressor on temperature deviation (δon), compressor off temperature deviation (δoff ) learned
for an AC usage. It also contains meta information such as AC usage duration, initial room
temperature and score, for every AC usage. PACMAN-Engine interacts with L-Records to get
thermal parameters for the prediction segment. Similarly, it makes an entry of learned thermal
parameters in L-Records for every usage.
P-Records contain metadata information such as initial external temperature (T 0ext), initial room
temperature (T 0int), AC set temperature (Tset) and a pointer to thermal parameters in L-Record
that were used for energy forecast. Prediction segment returns a vector of predicted room
temperature and compressor status using which PACMAN-Engine estimates energy forecast of
AC and minimum achievable temperature to store them in P-Records.
LP-Statistics keep information related to prediction error and accuracy. PACMAN-Engine an-
alyzes predicted and actual usage to generate statistics evaluating prediction performance of
PACMAN. Finally, it updates these statistics and various parameters of accuracy metrics in
LP-Statistics of PACMAN-Database.
21
Chapter 6
Evaluation
We conducted a real world study across nine different rooms in seven homes. The study was
started in August, 2013 and persisted for approximately three months. During our data collec-
tion, we collected room temperature, power consumption and external temperature data. Data
collection from each home was done in different time intervals across three months. Table-5.1
summarizes environment of the nine rooms, with their AC specifications, from which the data
was collected.
6.0.4 Experimental Setup
Figure 6.1: Installed Raspberry Pi for Data Collec-tion
A temperature sensor from Maxim1, with ac-
curacy of 0.5◦C, was interfaced with a single
board computer2 to record room temperature
data at 1Hz. Figure 6.1 presents labeled dia-
gram of a system used for data collection and
uploading it to the server. Collected data was
posted to a server over HTTP interface. For
each home, we collected power consumption
data from a modbus enabled electricity me-
ter with a sampling rate of 1Hz. Each home
had a three phase power supply with a clear
understanding of which phase was supplying power to the AC that we were monitoring. Cor-
respondingly, it was easy to extract the AC power consumption from the phase level power
consumption monitored at the meter level using simple transient based methodology [11]. Fig-
ure 6.2a illustrates AC power extraction from phase level power consumption from one of the
usage instance in one of the homes. External temperature was recorded using Wunderground
API3 that provided two readings of historical weather information for every hour and one read-
1http://bit.ly/1fdT62x2http://www.raspberrypi.org/3http://www.wunderground.com/
22
(a) AC power extraction from ag-gregate phase power measured atthe meter level
(b) Comparison of recorded exter-nal temperature on campus andpulled from weather service
(c) Comparison of weather fore-cast and historical weather as ex-tracted from weather service
Figure 6.2: Validation of different datasets used for the empirical evaluation of PACMAN.
ing of weather forecast every hour. During the course of study, we recorded average external
temperature to be 28◦C, with the minimum at 25◦C and maximum at 33◦C.
6.0.5 Data Validation
Datasets used as proxy, e.g. weather data from weather service as a proxy for external temper-
ature, are validated for their accuracy before using them for evaluation purpose. Figure 6.2a
validates power consumption extraction for an AC, using the meter data. During this experi-
ment, AC power consumption was separately monitored using jPlug [10]. AC power consumption
as extracted from meter data and as observed directly were observed to be similar.
To validate the accuracy of external temperature data as provided by weather service, we
recorded external temperature from our own temperature sensor for a day. For the same dura-
tion, historical weather temperature was also extracted from the weather service. Figure 6.2b
compares both the dataset confirming that the historical temperature provided by the weather
service is close to the actual external temperature as observed by us. The mean absolute devi-
ation4 of 0.35◦C was observed between the actual temperature and historical temperature from
the weather service.
Since PACMAN takes weather forecast, from the weather service, for the prediction of AC
energy consumption, the energy prediction accuracy directly depends on the accuracy of weather
forecast. To validate the accuracy of forecasts provided by the website, we collected both
the forecasted temperature and actual temperature for the duration of 34 hours, as shown in
Figure 6.2c. Mean absolute deviation4 of 0.83◦C was observed between the forecasted and actual
temperature, as reported by the weather service.
6.0.6 Analysis of AC Energy Prediction Accuracy
We now present the details of AC energy consumption prediction accuracy of PACMAN. We
evaluate the prediction accuracy of PACMAN for each room (with different combination of
room and AC specifications), each set temperature across different homes and ACs, and across
4http://bit.ly/TJzIm6
23
(a) Usage wise accuracy in energyprediction across all ACs at settemperature of 27◦C.
(b) Average compressor on and offtemperature for different AC settemperatures
(c) Variation in compressor on andoff temperature at set tempera-ture of 27◦C across different us-ages for Room-1.
Figure 6.3: Variation in compressor on and off temperatures across different set temperatures and usages.
different AC manufacturers to understand the general applicability of PACMAN across diverse
usages.
Room level comparison
Figure 6.4 presents box plot for room level prediction accuracy in energy forecast of AC usage
across all the 9 rooms that were studied. Maximum prediction accuracy across all the rooms was
observed to be more than 96%. Average prediction accuracy across all the rooms was observed
to be at least 70% except for the two rooms, Room-6 and Room-9, that showed poor prediction
accuracy for some of the usage instances. Overall prediction for Room-6 and Room-9 drops
because the thermal noise for these rooms is not accurately accounted for in our thermal model.
For example, Room-9 was a dining hall with a large open space resulting in significant thermal
noise and interference, leading to poor prediction accuracy for several usage instances. It had
minimum accuracy of 10% in AC energy consumption forecast.
AC Set Temperature Based Evaluation
During our experiment, users set AC at various set temperature, as shown in Column 3 of
Table-5.1. Figure 6.4b summarizes accuracy in energy prediction of AC usage based on different
set temperatures, for all the rooms. Across all the set temperatures, average and maximum
prediction accuracy was observed to be at least 66% and 87% respectively. There were only
a small number of usages at some of the set temperatures thus resulting in smaller variation
in prediction accuracy for such temperature settings. Set temperature of 18◦C was used only
for Room-8 and Room-9. Since prediction accuracy for Room-8 was consistently very high (as
can be observed in Figure 6.4), poor accuracy levels at 18◦C can be attributed to poor thermal
model learned for Room-9. Similar argument can also be extended for set temperatures of 18,
25, 26 and 27 ◦C as these set temperatures were used by Room-6 and Room-9, which gave
poor room level accuracy. Taking away these two rooms, prediction accuracy for PACMAN was
consistently observed to be more than 80% across all set temperatures.
24
(a) Comparison across differentrooms
(b) Comparison across differentset temperatures
(c) Comparison across differentmanufacturers
Figure 6.4: Comparison of AC Energy Prediction accuracy based on different influencing parameters -room (for thermal properties), set temperature (being directly used in the prediction model) and ACmanufacturer (possibly leading to different control mechanisms).
AC Manufacturer Based Evaluation
Figure 6.4c illustrates energy prediction accuracy based on variations in model of ACs used in
our experimentation. Room-9 and Room-6 had Voltas and Carrier air conditioners respectively
thus resulting in poor prediction accuracy for these manufacturers (mostly attributed to poor
thermal model learned for the room). Leaving these two manufacturers, PACMAN performs
reasonably well across different manufacturers with average and maximum prediction accuracy
to be at least 80% and 97% respectively.
Performance Improvement with Usage
PACMAN scoring based best fit model selection, as explained in Algorithm-2, should ideally
result in improved performance prediction with increasing number of usages. This was further
validated from the collected data as well. Figure 6.3a shows the prediction accuracy for different
usages at 27◦C, when averaged across all the homes. If, for one of the homes, AC was used at
27◦C only three times, it is counted in averaged data for the first three usages only. Maximum
number of usages at 27◦C across all the homes was 9. Improvement in prediction accuracy with
increasing usages validates the adaptive nature of PACMAN approach, making it better than
other static model based systems, proposed earlier [18].
Status Prediction Accuracy:
It denotes number of times status was predicted accurately. It is 100 minus mean absolute error
in predicting compressor status at each second. Figure-6.5a presents apartment wise box plot of
status prediction accuracy at each second. As compressor on and off temperature depends upon
thermal conditions around temperature sensor in the room, it was diffcult to model is precisely.
Therefore, status predition accuracy shows variations for each apartment. We will discuss about
compressor on and off temperature modelling later in this section.
25
(a) Room wise accuracy in statusprediction across all ACs.
(b) Room wise precision across allACs.
(c) Room wise recall across allACs.
Figure 6.5: Performance evaluation of PACMAN based on (a) Status prediction accuracy (b) Precision(c) Recall.
Precision and Recall:
Precision denotes number of times compressor status was predicted ”On” accuractely out of total
number of instances it was predicted to be ”On”. Figure-6.5b presents precision in predicting
compressor ”On” status for each apartment. Recall denotes number of predicted compressor
”On” instances to the number of actual compressor ”On” instances. Figure-6.5c shows recall
in predicting compressor ”On” status for each apartment. Due to low accuracy in predicting
compressor on and off status, preditcted and actual internal room temperature were not time
aligned thus reducing its accuracy, precision and recall. We will discuss about these challenges
later in the thesis.
26
Chapter 7
Discussion
PACMAN is a feedback system for the prediction of energy consumption and is principally
designed for room level ACs. It anticipates AC energy consumption followed by anatomization
of AC usage to guide users towards energy conscious utilization of AC and its optimal settings.
Prediction accuracy of PACMAN depends upon how well the thermal model can be learned which
in turn depends upon parameter selection and proximity of chosen parameters from present
thermal environment. We now analyse reasons for some of the outliers in prediction accuracy of
PACMAN, caused primarily from imperfect applicability of the proposed thermal model, thus
requiring inclusion of multiple thermodynamic properties to enrich the current model being
learned.
As is clear from Equation-5.2 and Equation-5.3, prediction accuracy of AC energy consumption
also depends on using correct compressor on and off temperature. These on/off temperatures
may vary based on different manufacturers and the location where the room temperature is
sensed for learning purposes. Currently, PACMAN predicts compressor on and off tempera-
tures statically, before it initiates room temperature prediction. However, thermal noise in the
room will impact the sensed room temperature and hence would result in varying compressor
on/off temperatures for different usages, as was also observed in our data collection. Figure 6.3b
presents average compressor on and off temperature, across all ACs, based on AC set temper-
ature. It shows that difference between compressor on and off temperature varies from 0.4◦C
to 5.13◦C depending upon the thermal environment in the room. Further, compressor off tem-
perature, that should ideally be close to AC set temperature, differs from set temperature by
0.44◦C to 1.93◦C across different set temperatures. Figure 6.3c presents usage wise variation in
compressor on and off temperature for Room-1 at set temperature of 27◦C. Large variation in
compressor on and off temperatures, for the same AC set temperature, motivates the need to
further enrich the proposed model.
Thermal noise, introduced by some unaccounted factors in the room, distorts temperature cycles
(resulting from compressor cycles) generated by the AC. Figure 7.1a presents an illustration of
one such scenario observed during our data collection. During the time durations 17:23 - 17:39
and 18:31 - 18:52, temperature variation in the room deviates from the normal cyclic pattern
27
(a) Thermal noise distorting cyclic pattern in roomtemperature from 17:23 - 17:39 and 18:31 - 18:52
(b) Phase change as external temperature becomeslower than room temperature.
Figure 7.1: Illustration of outlier usages resulting in reduced AC energy prediction accuracy
observed during other times. This deviation from cyclic pattern could be attributed to the
addition of some thermal noise (e.g. door/window being left open), thus making temperature
prediction (hence energy consumption) inaccurate. In the scenario, where external temperature
is lower than the AC set temperature, the compressor should ideally turn off. PACMAN pre-
diction model, accounting for variation in external temperature, will also work as per this ideal
scenario. However, if there are sources of additional thermal noise in the room, the compressor
may still continue running, thus resulting in poor prediction accuracy. Figure 7.1b illustrates
one such scenario observed during our data collection drive. Beyond 01:00 AM, external tem-
perature dropped lower than room temperature and correspondingly PACMAN model predicted
compressor to be off resulting in room temperature variation as per the thermal model. How-
ever, in reality, the compressor switching pattern continued (resulting in a cyclic pattern of
room temperature). As a result, AC energy consumption prediction accuracy for this usage was
reduced to 8%.
Variations in room temperature measurement depend upon numerous factors such as number of
occupants, human activities, number of doors & windows together with whether they are open or
closed and thermal appliances present in the room. However, annotating each factor would make
the model overly complex along with making it cumbersome for the user to accurately report each
of these parameters as and when they change. Therefore, while compromising on the accuracy
in some of the usage instances, we simplified the thermal model requiring only set temperature
to be reported by the user. Placement of temperature sensor in the room also impacts the
prediction accuracy as its temperature measurement would depend upon thermal conditions
surrounding the sensor. We deployed temperature sensor in some open space within the room
while also accounting for aesthetics in the room to ensure its presence is not discomforting to
the user.
28
Chapter 8
Realization of PACMAN
In this chapter, we explain realization of a prototype system using PACMAN for providing
feedback to the user towards optimal usage of AC. Our prototype is based on a wireless protocol
i.e. Z-Wave. Before looking into the architecture, we discuss, wireless technoogy Z-Wave followed
by working system architecture to control Room AC based on Z-Wave. We then conclude this
section with the implementation details of the proposed architecture.
8.1 Z-Wave Overview
Z-Wave is a low bandwidth half duplex wireless communication protocol which is designed for
home automation. Z-Wave uses a low power wireless technology, designed to control remote
applications. Z-Wave is primarily designed for low bandwidth communication and it works in
a frequency range of around 900 MHz. Z-Wave was designed by Danish startup called ZenSys
that was acquired by Sigma Designs in 2008. Z-Wave is a proprietary wireless protocol designed
specifically for residential control. Figure-8.1b represents network stack for Z-Wave which consist
of five layers. Each Z-Wave network may include up to 232 nodes and consist of controllers and
slave devices. Nodes are capable of retransmit and forward messages allowing multipath network
in an established network. Radio Specifications are as follows 1:
• Bandwidth: 9,600 bit/s or 40 kbit/s, fully interoperable
• Modulation: GFSK
• Range: Approximately 100 feet (or 30 meters) assuming ”open air” conditions, with
reduced range indoors depending on building materials
• Frequency band: The ZWave Radio uses the 868.42 MHz SRD Band (Europe) the
900 MHz ISM band: 908.42 MHz (United States) 919.82 MHz (Hong Kong) 921.42 MHz
(Australian/New Zealand).
1http://www.z-wave.com/
29
Z-Wave creates a mesh network and nodes are part of that network. Majorly, there are only two
type of nodes i.e. controller and slave nodes. Hierarchy for type of devices in a Z-Wave network
is as follows:
1. Controller Nodes: Controller Nodes are those Z-Wave devices that initiates commu-
nication between the devices. Commands are distributed in various command classes.
Commands are sent over Z-Wave network to slave devices to perform some action or re-
port about their current status. Controllers are of two types:
• Portable Controller: As the name suggest, portable controllers are designed to
change their position in the Z-Wave network. They are generally battery powered
devices and mainly used in portable applications. Example: Remote Controller
• Static Controller: Fixed controller that never changes its position. Z-Wave network
is generally created by a static controller that are also known as primary controller.
• Static Update Controller(SUC): Static controller receives notifications from the
primary controller regarding all the changes taking place in the network’s topology.
It is capable of sending network topology updates to other controllers and routing
slaves upon request.
2. Slaves: A Z-Wave device has very little information about the network and follow com-
mands sent by the controller. These devices are not having any capability to include/exclude
nodes to Z-Wave Network.
3. Home ID & Node ID: The Z-Wave protocol uses a unique identifier called Home ID
to separate networks from each other. Home ID is a 32-bit unique identifier which is
preprogrammed in all controller devices. Node IDs are used to address individual nodes
in the network. It is an 8-bit value assigned to slave nodes by the controller.
8.2 System Architecture
In this section, we discuss architecture of proposed system and its implementation details.
Figure 8.1a presents various components of the realized system together with the interaction
amongst them and with the user.
8.2.1 System Components
Our architecture includes two Z-Wave sensors, one to measure power consumption of AC and
other to control AC using the remote application (Ex. Web Interface, Mobile Application).
Some of the system components involved in the architecture are:
30
(a) System architecture of Z-Wave based AC ac-tuation and temperature monitoring system. Userinvolvement is through a web server hosted by thesingle board computer.
(b) Z-Wave Stack
Figure 8.1: System Description
1. Z-Wave Controller (HSC-48): It is a Z-Wave2 gateway manufactured by HomeSce-
nario3 that acts as a primary controller to create Z-Wave network. Within the system,
it functions as an interface between WiFi and ZWave networks. It supports a maximum
range of 30m indoor line of sight, sufficient for residential purpose.
2. Z-Wave Based Sensors:
(a) Power Monitor Switch: It is a Z-Wave based power module for monitoring power
consumption of AC and switching it on or off. It can work with a maximum load of
3000 Watts.
(b) HSK-200Z (IR Sensor): HSK-200Z, manufactured by HomeScenario, is an IR
sensor capable of emulating buttons of any IR based remote. It was used to provide
user with a capability to control their AC remotely, besides allowing the system to
automatically control the set temperatures.
8.2.2 System Implementation
A Single Board Computer (SBC) was used as a web server for hosting an application for user
interface. The web server provided used with access to real time AC power consumption, room
temperature and weather updates. Following is a sequence of communication between various
components that result in the realization of PACMAN architecture using the prototype system:
1. SBC hosts an application for the user, which provides an interface, to input the desired set
temperature for the AC and user’s location, for weather forecast information. Using this
2http://www.z-wave.com/3http://www.homescenario.com/home/index.html
31
Figure 8.2: User Interface to monitor AC power consumption and control AC remotely
information together with PACMAN learned model from past usages, a predicted energy
consumption at different set temperatures is displayed.
2. User can then select from any of the set temperatures and corresponding message is relayed
to the AC using HSK-200Z. Message relay occurs via HSC48 that takes the command on
WiFi and encodes it as a Z-Wave command before broadcasting it to the Z-Wave network,
for the end node.
3. As the AC is switched on, power monitor starts reporting its power consumption to SBC
via the HSC48 gateway. Accordingly, SBC provides real time power consumption together
with real time updates of room temperature (as reported by the attached temperature
sensor) and external temperature (as pulled from the weather service).
4. When AC is switched off (as reported by Power monitor or HSK-200Z), PACMAN uses
the temperature data from the past usage to update its model for future usages.
Web interface from current realization can easily be extended for mobile phones as well.
8.2.3 User Interface
We also developed an user interface to interact with Z-Wave based system and remotely send
commands to AC. Figure-8.2 presents screenshot of the user interface. Using this interface, user
can configure Z-Wave sensors, AC remote keys according to his comfort level. Plot provides
power consumption of AC in real time to the user. Power reading is reflected in the bottom
left corner of the interface from where user may actuate AC also. Log section in bottom right
corner of the interface is to display error messages and acknowledgments to the user.
32
Chapter 9
Conclusion
In this thesis, we present PACMAN system to forecast energy consumption for room level ACs.
The proposed system includes participatory involvement of the users by providing predicted en-
ergy consumption at different AC set temperatures thus motivating the user to set optimally the
AC temperature realizing both user comfort and energy efficiency. Extensive study, conducted
across nine rooms in seven different homes, validated the high prediction accuracy for PACMAN
system, across ACs with different manufacturers and with different set temperatures. A pro-
totype system implementation is presented, illustrating the simplicity of the hardware system
required for real world realization of PACMAN based AC control system. We outline different
scenarios in which the proposed system fails to predict the energy consumption accurately for an
AC usage, motivating the future work required towards enriching the thermal model proposed
in this thesis based on which AC consumption is predicted.
33
Bibliography
[1] Agarwal, Y., Balaji, B., Dutta, S., Gupta, R., and Weng, T. Duty-cycling build-
ings aggressively: The next frontier in hvac control. In Information Processing in Sensor
Networks (IPSN), 2011 10th International Conference on (April 2011), pp. 246–257.
[2] Agarwal, Y., Balaji, B., Gupta, R., Lyles, J., Wei, M., and Weng, T. Occupancy-
driven energy management for smart building automation. In In BuildSys (2010).
[3] Balaji, B., Xu, J., Nwokafor, A., Gupta, R., and Agarwal, Y. Sentinel: Occu-
pancy based hvac actuation using existing wifi infrastructure within commercial buildings.
In Proceedings of the 11th ACM Conference on Embedded Networked Sensor Systems (New
York, NY, USA, 2013), SenSys ’13, ACM, pp. 17:1–17:14.
[4] Batra, N., Arjunan, P., Singh, A., and Singh, P. Experiences with occupancy based
building management systems. In Intelligent Sensors, Sensor Networks and Information
Processing, 2013 IEEE Eighth International Conference on (2013), pp. 153–158.
[5] Batra, N., Kelly, J., Parson, O., Dutta, H., Knottenbelt, W., Rogers, A.,
Singh, A., and Srivastava, M. NILMTK: An Open Source Toolkit for Non-intrusive
Load Monitoring. In Fifth International Conference on Future Energy Systems (ACM e-
Energy) (Cambridge, UK, 2014).
[6] Clausius, R. X. on a modified form of the second fundamental theorem in the mechanical
theory of heat. Philosophical Magazine Series 4 12, 77 (1856), 81–98.
[7] Darby, S., et al. The effectiveness of feedback on energy consumption.
[8] Erickson, V., Carreira-Perpinan, M., and Cerpa, A. Observe: Occupancy-based
system for efficient reduction of hvac energy. In Information Processing in Sensor Networks
(IPSN), 2011 10th International Conference on (April 2011), pp. 258–269.
[9] Erickson, V. L., Achleitner, S., and Cerpa, A. E. Poem: Power-efficient occupancy-
based energy management system. In Proceedings of the 12th International Conference on
Information Processing in Sensor Networks (New York, NY, USA, 2013), IPSN ’13, ACM,
pp. 203–216.
34
[10] Ganu, T., Seetharam, D. P., Arya, V., Kunnath, R., Hazra, J., Husain, S. A.,
De Silva, L. C., and Kalyanaraman, S. nplug: A smart plug for alleviating peak
loads. In Proceedings of the 3rd International Conference on Future Energy Systems: Where
Energy, Computing and Communication Meet (New York, NY, USA, 2012), e-Energy ’12,
ACM, pp. 30:1–30:10.
[11] Hart, G. Nonintrusive appliance load monitoring. Proceedings of the IEEE 80, 12 (Dec
1992), 1870–1891.
[12] Jain, M., Chabra, D., Mankoff, J., and Singh, A. Energy Usage Attitudes of Urban
India. In Second International Conference ICT for Sustainability (Stockholm, Sweden,
2014).
[13] Li, B., and Alleyne, A. Optimal on-off control of an air conditioning and refrigeration
system. In American Control Conference (ACC), 2010 (June 2010), pp. 5892–5897.
[14] Lu, J., Sookoor, T., Srinivasan, V., Gao, G., Holben, B., Stankovic, J., Field,
E., and Whitehouse, K. The smart thermostat: Using occupancy sensors to save energy
in homes. In Proceedings of the 8th ACM Conference on Embedded Networked Sensor
Systems (New York, NY, USA, 2010), SenSys ’10, ACM, pp. 211–224.
[15] Nguyen, T. A., and Aiello, M. Energy intelligent buildings based on user activity: A
survey. Energy and Buildings 56, 0 (2013), 244 – 257.
[16] Prez-Lombard, L., Ortiz, J., and Pout, C. A review on buildings energy consumption
information. Energy and Buildings 40, 3 (2008), 394 – 398.
[17] Rajput, R. K. Engineering thermodynamics. Jones and Bartlett Publishers, Sudbury,
Mass, 2010, ch. REFRIGERATION CYCLES.
[18] Rogers, A., Ghosh, S., Wilcock, R., and Jennings, N. R. A scalable low-cost
solution to provide personalised home heating advice to households. In Proceedings of the
5th ACM Workshop on Embedded Systems For Energy-Efficient Buildings (New York, NY,
USA, 2013), BuildSys’13, ACM, pp. 1:1–1:8.
[19] Rogers, A., Maleki, S., Ghosh, S., and Nicholas R, J. Adaptive home heating
control through gaussian process prediction and mathematical programming. In Second
International Workshop on Agent Technology for Energy Systems (ATES 2011) (May 2011),
pp. 71–78. Event Dates: May 2011.
[20] Scott, J., Bernheim Brush, A., Krumm, J., Meyers, B., Hazas, M., Hodges,
S., and Villar, N. Preheat: Controlling home heating using occupancy prediction. In
Proceedings of the 13th International Conference on Ubiquitous Computing (New York, NY,
USA, 2011), UbiComp ’11, ACM, pp. 281–290.
35
[21] Weng, T., Balaji, B., Dutta, S., Gupta, R., and Agarwal, Y. Managing plug-
loads for demand response within buildings. In Proceedings of the Third ACM Workshop
on Embedded Sensing Systems for Energy-Efficiency in Buildings (New York, NY, USA,
2011), BuildSys ’11, ACM, pp. 13–18.
36